Hark, Brett Adcock’s (founder of Figure AI) consumer AI lab, has released Handoff² — the second-generation computer-use agent that controls the browser and completes tasks: buying groceries, booking tickets, and scheduling doctor appointments. According to the company, the agent is on average twice as fast as the first version, and the update is already rolling out to all Hark Pro plans.

image
image

What happened

The release covers both the model and the surrounding environment. The Handoff² model itself is, according to Hark, 46% faster than the first version, while the company attributed the rest of the speedup to backend improvements. Specific scenarios saw even greater gains: airline ticket booking was up to 3.8 times faster, and long research tasks were more than 5 times faster. On an internal benchmark of 478 tasks across 238 sites, the success rate rose from 73% to 82%. On an internal behavior test, which includes checks prohibiting purchases without user consent, the result improved from 90.3% to 93.8%.

Context

The first version of Handoff launched in October, and since then the agent has conducted over 5.6 million browser sessions and processed purchases totaling more than $20 million, meaning autonomous transactions are already happening at a significant scale. The main signal of the release is how the speedup was achieved. The primary contribution came from the joint development of the harness and RL recipe, which penalizes extra moves, excessive thinking tokens, and latency: the agent learns to batch obvious actions, such as filling multiple form fields in one move, and to collapse repeating sequences into reusable composite actions. According to Hark, Handoff² is within 5 points of Claude Opus 5.5 (max) with half the latency, although this comparison was conducted in Hark’s harness, not in a neutral environment.

Why this matters for the industry

For the industry, the key aspect of the release is that agent speed is shown as a learnable metric, not a side effect of model size. If the recipe with penalties for extra moves and latency is reproducible, the unit economics of computer-use agents can be improved without moving to a larger model: fewer moves and thinking tokens mean a lower cost per task. Reusable composite actions are likely to become a standard feature in browser-use stacks, and competitors will presumably respond with their own harness optimizations, making a latency race and agent comparisons based on price and task completion time—rather than just success rates—likely. For external teams, the product remains closed: there is no public API, the recipe cannot be applied directly, and only general principles are available.

Why this matters for users

Handoff² is already rolling out to all Hark Pro plans, so users get the speedup without changing plans: entry is free, and paid plans at $20 and $100 per month simply increase usage limits. The company estimates the effect as follows: a 10-minute task shrinks to 3 minutes, and a 45-minute task to 10. In practice, this applies to common scenarios like booking tickets or placing orders, which the agent completes, including payment. Since the agent processes payments itself, it is still worth reviewing the outcome of large purchases or bookings, especially for high amounts.

What is still unknown / limitations

All figures are Hark’s internal benchmarks, not independently verified: the task distribution, confidence intervals, and whether the task sets for the two agent versions were identical are unknown. There is no neutral comparison with flagship models, so the correct conclusion is limited to Hark’s own stack: their agent is close to the flagship on their metric with lower latency. Only general principles of acceleration have been disclosed: reward functions and their weights, the action space for batching, criteria for selecting obvious actions, and the details of backend improvements are not described, and there are no articles or tech reports with methodology. Finally, even after the improvement, about 6% of behavior checks still fail, and at the scale of already running autonomous sessions, rare transaction failures stop being statistics and become an operational risk.

Sources

Author

Look at AI, editorial team